Optimal features assisted multi-attention fusion for robust fire recognition in adverse conditions

Area of Science:

• Computer Vision
• Artificial Intelligence
• Fire Safety Engineering

Background:

• Deep neural networks show promise in visual fire detection but suffer from high false alarm rates and poor performance in challenging environments.
• Existing methods often utilize shallow networks, limiting their ability to capture complex fire features and context.

Purpose of the Study:

• To develop a novel deep learning framework, the Attention-Enhanced Fire Recognition Network (AEFRN), to overcome the limitations of current visual fire detection systems.
• To improve accuracy and reduce false alarms in fire detection while maintaining computational efficiency for practical applications.

Main Methods:

• Introduced Convolutional Self-Attention (CSA) integrating self-attention with convolution for enhanced low-level feature processing.
• Developed Recursive Atrous Self-Attention (RASA) to capture multi-scale contextual information efficiently.
• Enhanced the Convolutional Block Attention Module (CBAM) for robust feature discrimination.

Main Results:

• Achieved state-of-the-art (SOTA) performance on the FD dataset with 99.11% accuracy and the BoWFire dataset with 97.98% accuracy.
• Demonstrated superior performance compared to twelve SOTA methods in challenging fire detection scenarios.
• Validated model interpretability using Grad-CAM, showing focus on fire-relevant regions.

Conclusions:

• AEFRN effectively addresses limitations in visual fire detection, offering high accuracy and reliability.
• The proposed attention mechanisms enhance feature processing and contextual understanding for robust fire recognition.
• AEFRN provides a computationally efficient and practical solution for advanced fire detection systems.